[Mlir-commits] [mlir] [llvm] [mlir] Lower math dialect later in gpu-lower-to-nvvm-pipeline (PR #78556)

[Adam Paszke]

================
@@ -0,0 +1,18 @@
+// RUN: mlir-opt -gpu-lower-to-nvvm-pipeline="cubin-format=isa" -split-input-file %s | FileCheck %s
+
+// CHECK-LABEL: llvm.func @test_math(%arg0: f32) {
+func.func @test_math(%arg0 : f32) {
+    %c2 = arith.constant 2 : index
+    %c1 = arith.constant 1 : index
+    // CHECK: gpu.launch_func  @test_math_kernel::@test_math_kernel
+    // CHECK: gpu.binary @test_math_kernel  [#gpu.object<#nvvm.target
+    gpu.launch 
+        blocks(%0, %1, %2) in (%3 = %c1, %4 = %c1, %5 = %c1) 
+        threads(%6, %7, %8) in (%9 = %c2, %10 = %c1, %11 = %c1) { 
+         %s1 = math.exp %arg0 : f32
----------------
apaszke wrote:

Add a check that there is a call to the libdevice function?

https://github.com/llvm/llvm-project/pull/78556